Pith. sign in

REVIEW 3 major objections 6 minor 49 references

A single-call prompt that asks language models to write the governing equations before solving outperforms chain-of-thought by 5.76 percentage points and program-of-thought by 8.42 on average across four applied-math benchmarks, with the la

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

An equation-first prompt beat chain-of-thought and program-of-thought by 5.76 and 8.42 points on average across four applied-math benchmarks, but the body does not include several analyses the abstract promises.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection The prompt works on the reported numbers, but the paper doesn't establish that equation formalization is the driver—the two conditions differ in more than Phase 1, and the largest gain sits on the most confounded benchmark. the 3 major comments →

arxiv 2601.19302 v3 pith:IRG6TQ6A submitted 2026-01-27 cs.CL

Formula-One Prompting: A Composable Equation-First Prefix for Applied Mathematics

Natapong Nitarach , Pittawat Taveekitworachai , Kunat Pipatanakul This is my paper
classification cs.CL
keywords formula-one promptingequation formalizationchain-of-thoughtprogram-of-thoughtapplied mathematicsLLM mathematical reasoningsingle-call prompting
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper aims to establish that language models already know the governing equations of applied math from pretraining, but standard prompts never elicit them, leaving useful knowledge dormant. It proposes Formula-One Prompting (F-1), a single-call prompt that makes the model identify givens, write the key equations, and only then choose a solving style—direct calculation, step-by-step reasoning, or code—with no explicit routing rules. Across five models and four applied-math benchmarks, F-1 beats chain-of-thought by 5.76 percentage points on average and program-of-thought by 8.42, with the largest gain on finance problems. Per-problem flip analysis leads the authors to attribute the gain primarily to the equation-formalization step rather than the strategy menu.

Core claim

This paper claims that an equation-first prompt fills what it calls the 'equation gap': LLMs encode governing equations from pretraining, but neither chain-of-thought nor program-of-thought prompts ask for them. F-1 inserts a formalization phase in which the model states givens and targets and writes the governing equations, followed by a natural-flow phase in which the model freely solves via direct calculation, CoT, PoT, or elimination, then verifies. In a single greedy-decoding call across 2,116 finance, physics, cryptography, and competition-math problems, F-1 beats CoT by 5.76 points and PoT by 8.42 points on average, winning 53 of 60 benchmark-model comparisons; the largest gain is +13

What carries the argument

The central object is the F-1 prompt itself: a unified system-and-user prompt that forces a two-phase response in one call. Phase 1 (Equation Formalization) asks the model to identify givens, identify the target, and write the key equations; Phase 2 (Natural Flow Reasoning) lets the model pick Direct, CoT, PoT, or elimination based on the equation structure, followed by a verification step. The load-bearing mechanism is Phase 1: the paper argues that writing equations up front exposes mathematical structure, prevents semantic confusions, guides format choices, and makes the model self-check against the equations.

Load-bearing premise

The causal claim that equation formalization drives the gains rests on the premise that the CoT and F-1 conditions differ only in the formalization phase; Appendix A shows they also differ in system prompts, the Phase-2 strategy menu, the verification instruction, and benchmark-specific output contracts, so the per-problem flip analysis may not isolate formalization.

What would settle it

Run a controlled ablation on the same five models and four benchmarks in which the F-1 prompt is modified to keep every element—system prompt, strategy menu, verification step, output contract—except Phase 1, replacing 'write the key equations' with 'restate the givens and target in plain language.' If accuracy stays at F-1 levels, the primary-driver claim collapses. A cheaper check: give chain-of-thought the same output contract and verification instruction on the finance benchmark and see whether the +13.30-point gap disappears.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Equation elicitation is a cheap, single-call intervention: it adds only about 68 input tokens over zero-shot prompting while improving accuracy across five model families and sizes.
  • Applied domains benefit more than pure competition math, suggesting equation-first prompting is most valuable where domain constraints are explicit in governing equations.
  • The model's solving strategy adapts to the problem without routing rules: computation-heavy problems draw more code-style solving, while proof-heavy problems draw more step-by-step and hybrid responses.
  • The benefit scales down to at least a 30B-parameter model, which gains about 5.6 points over chain-of-thought, comparable to frontier-model gains.
  • Most remaining failures happen after a correct Phase-1 equation, so the next bottleneck is in the solving or verification phase, not formalization.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the appendix shows the CoT and F-1 prompt conditions differ in more than Phase 1 (system prompts, strategy menu, verification instruction, and per-benchmark output contracts), a strict reader should treat the 'formalization is the driver' conclusion as not fully isolated; the flip analysis measures the whole prompt package.
  • This reading suggests a direct test: hold every other prompt element fixed and replace 'write the key equations' with an instruction to restate the givens and target in plain language; if the gain persists, formalization per se is not the active ingredient.
  • The equation-gap idea points to other structured expert notations that pretraining may also encode but prompts ignore—e.g., chemical reaction schemes, circuit equations, or statistical model specifications—where a similar formalize-first prompt might transfer.
  • The authors' own speculation about training-data composition could be tested by ablating the ratio of equation-rich text to step-by-step narration in pretraining or fine-tuning corpora; the paper leaves this to future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper introduces Formula-One Prompting (F-1), a single-call, two-phase prompting method for applied mathematics: Phase 1 asks the model to identify givens/targets and write governing equations; Phase 2 lets the model choose among CoT, PoT, direct calculation, or elimination, followed by verification. The authors motivate this with a corpus statistic that equation-centered language is more frequent in pretraining data than code or step-by-step narration. They evaluate F-1 against Zero-Shot, CoT, and PoT on IMO-Bench, OlympiadBench, FinanceMath, and AICrypto across five models, reporting average gains of +5.76 pp over CoT and +8.42 pp over PoT, with the largest gain (+13.30 pp) on FinanceMath, and 53/60 wins. Section 6 attempts to attribute the gain to equation formalization via per-problem flip analysis and strategy-selection analysis. The full prompt templates are given in Appendix A.

Significance. If the causal claim holds, F-1 would be an unusually cheap and broadly applicable intervention: a single call, about 68 extra input tokens, no fitted parameters, no external routing, and consistent gains across model families, with particularly large effects in finance and physics. The strengths of the paper are real: evaluation on public benchmarks, a mix of deterministic regex grading and LLM judges, full prompts in the appendix, five models from 30B to frontier scale, and a token-efficiency analysis. The headline accuracy comparison of the complete F-1 prompt against the complete CoT/PoT prompts is a legitimate method-level result. However, the paper's stronger claim—that equation formalization is the primary driver—currently rests on a false premise about the prompt designs. The confound is serious enough that the causal attribution and the interpretation of the FinanceMath gain are not yet established, though they are fixable with targeted ablations.

major comments (3)
  1. [§6.1, Appendix A] The per-problem flip analysis is built on the premise "CoT and F-1 differ only in the equation formalization phase (Phase 1)." Appendix A shows this premise is false. The F-1 condition differs from CoT in at least four ways beyond Phase 1: (a) a different system prompt ('You are an AI assistant that solves problems mainly through equations' vs. 'think step-by-step'); (b) a Phase-2 strategy menu explicitly naming CoT, PoT, direct calculation, and elimination; (c) a verification instruction; and (d) benchmark-specific output contracts, most prominently FinanceMath's 'OUTPUT CONTRACT (strict): The very last line MUST be exactly: Final Answer (3 decimal)' plus the extra user prefix 'Write equations (LaTeX) with minimal text'. Table 6 therefore measures the combined effect of all these differences, not the isolated contribution of equation formalization. An ablation that removes Phase 1 while
  2. [§5.1, Figure 3, Appendix D.2] The FinanceMath result is the largest and most striking gain (+13.30 pp over CoT), but the paper's own qualitative example shows that format, not equation formalization, accounts for at least some of it. In Appendix D.2, Zero-Shot computes 6.3%, CoT computes 6.3%, and PoT computes 6.252%; all are marked incorrect because the expected answer is the decimal 0.063. F-1's prompt, unlike CoT's, contains a strict output contract for the final answer and a user prefix 'Write equations (LaTeX) with minimal text; show steps clearly.' The per-problem flip analysis in Table 6 excludes FinanceMath entirely, so the benchmark with the largest reported gain is never subjected to the claimed isolation. A format-matched control—e.g., CoT with the same output contract but no Phase 1—is necessary to separate format compliance from equation-formalization effects. Without it, the FinanceMath contribution to
  3. [§6.3, Table 2, Appendix E.1] The strategy-selection analysis is presented as evidence for the 'natural flow reasoning' mechanism, but the methodology is inconsistent. Table 2 states that in F-1, PoT 'uses code as a reasoning format; the model traces through code mentally without execution,' while Appendix E.1 states that 'When F-1 selects PoT as its solving strategy, the generated code is executed identically to the baseline' in a sandbox. These cannot both be true. If code is executed, then F-1 is not purely single-call in its reasoning (though it remains single-call in generation), and the 'mental tracing' description in Table 2 and Figure 1 is inaccurate. If code is not executed, the evaluation protocol for F-1's PoT path is unclear. This ambiguity affects the interpretation of the strategy distributions in Table 8 and the '81–84% of upper bound' claim. The authors should state clearly whether F-1's PoT outputs a
minor comments (6)
  1. [Table 12 vs Table 4] Table 12 says GPT-5 data are unavailable for AICrypto token efficiency, yet Table 4 reports GPT-5 AICrypto accuracy (98.50). Please clarify whether the accuracy is from a different run or whether the token data are missing.
  2. [Table 4 and Table 23] The abstract and Table 4 report F-1 overall average as 61.06, while Table 23 reports 61.07. Please make the numbers consistent.
  3. [Abstract/§1] The infini-gram-mini corpus statistic (121x vs code, 3.79x vs step-by-step) is mentioned in the abstract as part of the motivation but is not described or cited in the main text. Either add a short methods sentence or remove the statistic from the abstract.
  4. [§4.2/Appendix E] The phrase 'rule-based scoring with deterministic regex-based extraction' should be reconciled with Appendix E.1's IMO-AnswerBench auto-grader, which is an LLM prompt that asks for mathematical equivalence. If an LLM is used for IMO-AnswerBench, it is not a pure regex, and this should be stated in §4.2.
  5. [Table 21] The 'Upper Bound' definition (100% minus 'All Failed') is the best any single baseline achieves, not the true upper bound across methods. Consider renaming it 'Best Single Baseline' to avoid overstatement, since F-1 can and does exceed it in some categories.
  6. [Limitations] The Limitations section appropriately notes the small sizes of AICrypto (n=18) and OlympiadBench TP_physics (n=25). Given the paper's emphasis on applied domains, a sentence in the main results cautioning that the AICrypto and TP_physics gains rest on very small samples would be appropriate.

Circularity Check

0 steps flagged

No circularity found: evaluation is external and parameter-free; the Section 6.1 attribution confound is a validity issue, not a circular reduction.

full rationale

The paper's accuracy claims are empirical measurements on external public benchmarks (IMO-Bench, OlympiadBench, FinanceMath, AICrypto) using deterministic regex grading and LLM judges. No parameters are fitted, and no prediction is defined in terms of a fitted quantity or of the reported result. F-1's gains are observed comparisons against Zero-Shot, CoT, and PoT baselines, not consequences of the prompt by construction. There are no load-bearing self-citations: the reference list contains no prior work by the present authors, and no uniqueness theorem is imported from the authors' own prior work. The per-problem flip analysis in Section 6.1 relies on the premise that 'CoT and F-1 differ only in the equation formalization phase (Phase 1)', but Appendix A shows additional prompt differences (system prompt, Phase-2 strategy menu, verification instruction, and a FinanceMath output contract), so the causal attribution to Phase 1 is confounded. That is a soundness/control problem rather than circularity: the quoted gain numbers are still externally measured and are not entailed by the prompt design itself. Accordingly, the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

No numerical parameters are fitted: the paper is an empirical evaluation, not a derivation. The load-bearing assumptions concern experimental isolation, judge validity, small-sample representativeness, and greedy-decoding stability. The abstract's corpus-scan ratios (121x, 3.79x) are asserted without methodology and are therefore unverifiable, not fitted parameters.

axioms (5)
  • ad hoc to paper CoT and F-1 differ only in the equation-formalization phase (Phase 1).
    Invoked in Section 6.1 but contradicted by Appendix A: system prompts, strategy menu, verification, and output contracts differ.
  • domain assumption LLM-as-judge scores are valid ground truth for proof-based problems.
    Section 4.2/E.2 uses Gemini-3-Pro and GPT-5.1 judges; judge reliability is not validated on a labeled subset.
  • domain assumption Small benchmark subsets are representative: AICrypto n=18, TP_physics n=25.
    Section 4.1 and Limitations acknowledge small N; conclusions on applied domains lean on these subsets.
  • domain assumption Single-sample greedy decoding at temperature 0 gives stable method comparisons.
    Section 4.2; no variance or multiple-seed evaluation is reported.
  • domain assumption The selected benchmarks (excluding GSM8K/MATH) are unsaturated and equation-driven.
    Section 4.1 criteria; the domain hypothesis depends on this selection being representative of applied-math reasoning.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Formula-One Prompting: A Composable Equation-First Prefix for Applied Mathematics." pith.science (2026). https://pith.science/paper/IRG6TQ6A

@misc{pith2026260119302,
  author       = {Pith},
  title        = {Pith review of: Formula-One Prompting: A Composable Equation-First Prefix for Applied Mathematics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IRG6TQ6A}},
  note         = {Machine review of arXiv:2601.19302}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This paper introduces Formula Prompting (FP) and Formula-One Prompting (F-1), two single-call methods that elicit governing equations before solving applied-math problems. Chain-of-Thought (CoT) and Program-of-Thought (PoT) prompting improve mathematical reasoning by eliciting reasoning traces or code-like structures learned during pretraining. This suggests a diagnostic question: which useful pretraining patterns remain under-elicited? Using infini-gram-mini, we scan 81.7 trillion pretraining tokens and find that, in curated corpora such as DataComp-LM, equation-centered language appears 121x more often than code and 3.79x more often than step-by-step narration, yet standard prompting methods do not explicitly elicit equation formulation. FP asks the model to formalize a problem's governing equations before solving; F-1 extends FP with a composable Phase 2 that selects Direct, CoT, or PoT-style solving in the same call. Across five reasoning models and four applied-math benchmarks (finance, physics, cryptography, competition math), F-1 outperforms CoT by 5.76 pp and PoT by 8.42 pp on average, with the largest gain of 13.30 pp on FinanceMath, while topping the accuracy-token efficiency frontier at only 68 prompt tokens of overhead. Variant ablations identify the equation-formalization prefix, not the strategy menu, as the primary driver: adding CoT or PoT on top of the prefix yields no further gain, and 73.3% of remaining failures occur downstream of a correct Phase-1 equation.

Figures

Figures reproduced from arXiv: 2601.19302 by Kunat Pipatanakul, Natapong Nitarach, Pittawat Taveekitworachai.

Figure 1
Figure 1. Figure 1: Comparison of prompting methods for mathematical reasoning. Traditional methods (Zero-shot, CoT, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: F-1 strategy distribution across domains [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison. (Left) Physics: Baselines confused “independent of x” with “equals zero”; F-1’s formalization prevented this. (Center) Finance: Only F-1 output correct decimal format (0.063). (Right) Cryptography: F-1 identified correct sub-hybrid argument while baselines used invalid reductions. Limitations Our evaluation spans models from 30B to frontier scale; we do not include smaller models (e… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 6 linked inside Pith

  1. [1]

    First, carefully read and understand the problem

  2. [2]

    Identify what is given and what needs to be found

  3. [3]

    Break down the problem into smaller steps

  4. [4]

    Preprint, arXiv:2410.05229

    Gsm-symbolic: Understanding the limitations of mathematical reasoning in large language models. Preprint, arXiv:2410.05229. Ivan Moshkov, Darragh Hanley, Ivan Sorokin, Shub- ham Toshniwal, Christof Henkel, Benedikt Schif- ferer, Wei Du, and Igor Gitman. 2025. Aimo-2 win- ning solution: Building state-of-the-art mathematical reasoning models with openmathr...

    Pith/arXiv arXiv 2025

  5. [5]

    Zayne Sprague, Fangcong Yin, Juan Diego Rodriguez, Dongwei Jiang, Manya Wadhwa, Prasann Singhal, Xinyu Zhao, Xi Ye, Kyle Mahowald, and Greg Durrett

    Llm-sr: Scientific equation discovery via pro- gramming with large language models.Preprint, arXiv:2404.18400. Zayne Sprague, Fangcong Yin, Juan Diego Rodriguez, Dongwei Jiang, Manya Wadhwa, Prasann Singhal, Xinyu Zhao, Xi Ye, Kyle Mahowald, and Greg Durrett. 2025. To cot or not to cot? chain-of- thought helps mainly on math and symbolic reason- ing.Prepr...

    Pith/arXiv arXiv 2025

  6. [6]

    Preprint, arXiv:2307.10635

    Scibench: Evaluating college-level scientific problem-solving abilities of large language models. Preprint, arXiv:2307.10635. 9 Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdh- ery, and Denny Zhou. 2023b. Self-consistency im- proves chain of thought reasoning in language mod- els.Preprint, arXiv:2203.11171. Yu Wan...

    Pith/arXiv arXiv 2025

  7. [7]

    So the final answer is {boxed_format}

    Formal mathematical reasoning: A new fron- tier in ai.Preprint, arXiv:2412.16075. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. 2023a. Tree of thoughts: Deliber- ate problem solving with large language models. Preprint, arXiv:2305.10601. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. ...

    Pith/arXiv arXiv 2024

  8. [11]

    Solve each step systematically, showing your reasoning

  9. [12]

    So the final answer is {boxed_format}

    Combine the results of each step to arrive at the final answer Please end your solution with "So the final answer is {boxed_format}." PoT User {problem_statement} {answer_type_text}Please solve this problem by:

  10. [13]

    Writing Python code to solve the problem programmatically

  11. [14]

    Showing your code and explaining each step

  12. [15]

    F-1 User (Ours) {problem_statement} {answer_type_text} Please solve using an Equation-First approach:

    Running the code to get the numerical answer You can use libraries like math, numpy, sympy, etc. F-1 User (Ours) {problem_statement} {answer_type_text} Please solve using an Equation-First approach:

  13. [16]

    List key variables and state the target quantity

  14. [17]

    Write the main governing equations/identities in LaTeX first

  15. [18]

    - PoT: write Python code to implement the equations

    After the equations, choose any suitable solving style: - CoT: think step-by-step, solve systematically, then combine results. - PoT: write Python code to implement the equations. - Zero-Shot: derive directly from equations with minimal text

  16. [19]

    So the final answer is {boxed_format}

    Combine the results of each step to arrive at the final answer Please end your solution with "So the final answer is {boxed_format}." A.3 FinanceMath Prompts A.3.1 System Prompts Zero-Shot System You are a financial expert, you are supposed to answer the given question. Therefore, the answer is {final answer}. The final answer should be a numeric value (3...

  17. [20]

    This mirrors expert problem-solving behavior and reduces ambigu- ity

    Equation-First Formalization: Problems are first translated into mathematical equations be- fore any solving attempts. This mirrors expert problem-solving behavior and reduces ambigu- ity

  18. [21]

    No explicit routing rules are imposed; the model freely rea- sons from equations to solution

    Natural Flow Reasoning: After formalization, the model naturally flows into the most appropri- ate solving method (CoT, PoT, or direct calcu- lation) based on equation structure. No explicit routing rules are imposed; the model freely rea- sons from equations to solution

  19. [22]

    B Token Efficiency Analysis This appendix provides detailed token efficiency analysis across models and benchmarks

    Verification Phase: Solutions are checked against the original equations, providing a built- in error-detection mechanism that catches com- putational and logical errors. B Token Efficiency Analysis This appendix provides detailed token efficiency analysis across models and benchmarks. We use two complementary metrics following Chen et al. (2023a) and Lee...

    arXiv 2025

  20. [23]

    Understand the problem: We are given general transformations x′ =X(x, t, v), t′ =T(x, t, v)between inertial frames and asked to use homogeneity of space and time to show that ∂X/∂x is independent ofxand∂T /∂tis independent oft

  21. [24]

    Need to show: ∂X/∂x does not depend on x; ∂T /∂t does not depend ont

    Given and to find: Given: Homogeneity of space and time (invariance under translations ofxandt). Need to show: ∂X/∂x does not depend on x; ∂T /∂t does not depend ont

  22. [25]

    Use this to relate X(x+a, t+b, v) to X(x, t, v)up to a constant

    Break into steps: Express homogeneity via invariance under shifts: x→ x+a , t→t+b . Use this to relate X(x+a, t+b, v) to X(x, t, v)up to a constant. Differentiate to see how partial derivatives behave. Conclude independence

  23. [26]

    x: ∂X/∂x(x+a, t+ b, v) =∂X/∂x(x, t, v)

    Solve systematically: By homogeneity, shifting the origin by (a, b)changes transformed coordinates only by an additive constant: X(x+a, t+b, v) =X(x, t, v) +Cx(a, b, v) T(x+a, t+b, v) =T(x, t, v) +Ct(a, b, v) Differentiate first equation w.r.t. x: ∂X/∂x(x+a, t+ b, v) =∂X/∂x(x, t, v). Since a, barbitrary, ∂X/∂x is invariant under shifts ⇒ cannot depend onx...

  24. [27]

    dX/dx =", dXdx) print(

    Combine results: The homogeneity implies the x-dependence of ∂X/∂x and t-dependence of ∂T /∂tvanish. Numerically: ∂/∂x(∂X/∂x) = 0and∂/∂t(∂T /∂t) = 0. So the final answer is \boxed{0}. CoT Judge Feedback Feedback:You correctly use homogeneity (translation invariance) and the idea of shifting coordinates to ar- gue that the partial derivatives are invariant...

  25. [28]

    Numerically equivalent to showing ∂/∂x(∂X/∂x) = 0and∂/∂t(∂T /∂t) = 0

    Key variables and target: Variables:x, t, v Target: ∂X/∂x independence of x and ∂T /∂tinde- pendence of t. Numerically equivalent to showing ∂/∂x(∂X/∂x) = 0and∂/∂t(∂T /∂t) = 0

  26. [29]

    T(x, t+ ∆, v)−T(x, t, v)depends only on ∆ and v (not onxort)

    Main equations: Definition of partial derivative: ∂X/∂x(x, t, v) = lim h→0 [X(x+h, t, v)−X(x, t, v)]/h ∂T /∂t(x, t, v) = lim k→0 [T(x, t+k, v)−T(x, t, v)]/k Homogeneity (translation invariance): X(x+ ∆, t, v)−X(x, t, v)depends only on ∆ and v (not onxort). T(x, t+ ∆, v)−T(x, t, v)depends only on ∆ and v (not onxort)

  27. [30]

    Hence ∂X/∂x(x, t, v) = limh→0 f(h, v)/h=g(v) , independent ofx

    CoT: By homogeneity, X(x+h, t, v)−X(x, t, v) =f(h, v), independent ofxandt. Hence ∂X/∂x(x, t, v) = limh→0 f(h, v)/h=g(v) , independent ofx. Therefore∂/∂x(∂X/∂x) = 0. Similarly,∂T /∂t=r(v), so∂/∂t(∂T /∂t) = 0

  28. [31]

    Correct Calculation, Wrong Format

    Conclusion: The spatial derivative ∂X/∂x is independent of x; the time derivative ∂T /∂tis independent of t. Numerically: ∂/∂x(∂X/∂x) = 0and∂/∂t(∂T /∂t) = 0. So the final answer is \boxed{0}. F-1 Judge Feedback Feedback:Your use of homogeneity via translation invariance and the limit definition of partial deriva- tives is solid and essentially complete. T...

  29. [32]

    Problem Formalization Model:Grinold–Kroner expected equity return Givens:Income return = 2.4%, Real growth = 5.0%, In- flation = 2.3%, P/E:14.5→14.0 Target:Expected annual returnas decimal (not percent- age)

  30. [33]

    Governing Equation E(R) = D P +g nominal + ∆(P/E) P/E0

  31. [34]

    Step-by-Step Calculation (in decimal) Dividend yield:0.024 Nominal growth:0.05 + 0.023 = 0.073 P/E change:(14.0−14.5)/14.5 =−0.034482759 Total:0.024 + 0.073−0.034482759 = 0.062517241

  32. [35]

    Correct Calculation, Wrong For- mat

    Format Verification Round to 3 decimals:0.062517...→0.063 Error Pattern: “Correct Calculation, Wrong For- mat” Observation:All methods computed ≈6.3% but only F-1 output correct format. Method Computed Output Issue Zero-Shot 6.3% 6.3 % instead of decimal CoT 6.3% 6.3 % instead of decimal PoT 6.252% 6.252 *100conversion F-16.3%0.063Correct format Root Caus...

  33. [36]

    Goal & Security Assumptions ProveAdv PRF H,A(n)is negligible for any PPTA

  34. [37]

    Hybrid Argument Structure Game 0:OracleO 0(x) =f n gn j (x)(s)Game 1:Oracle O1(x) =f n rG(x)(s)Game 2:OracleO 2(x) =r(x)

  35. [38]

    Game 0 → Game 1:Standard reduction to G’s PRF security.|P 0 −P 1| ≤negl(n)✓

  36. [39]

    If F is a PRF , then for any fixed s, the distribution {f n k (s)}k←{0,1}n is computationally indistinguishable from uniform

    Key Lemma for Game 1→Game 2: “If F is a PRF , then for any fixed s, the distribution {f n k (s)}k←{0,1}n is computationally indistinguishable from uniform. ” Proof:Build BF that queries its oracle at single point s. If oracle is f n k , output is from {f n k (s)}; if random function, output is uniform

  37. [40]

    varying keys, fixed input

    Sub-Hybrid Argument:Define H0, . . . , Hq where Hi answers first i queries with random values. Each |P(H i−1)−P(H i)| bounded by F ’s security. Sum over qsteps remains negligible. Conclusion: |P0 −P 2| ≤ |P0 −P 1|+|P 1 −P 2| ≤ negl(n)∴His a PRF family.□ F-1 Judge Feedback This is anexcellent, textbook-quality proof. The hybrid argument is clearly structur...

    2025

  38. [41]

    Analyze problem and ground truth solution

  39. [42]

    Step-by-step verification of every logical step

  40. [43]

    Identify all flaws, gaps, and errors

  41. [44]

    Compare against grading guidelines Output:<points>N out of 7</points> AICrypto (CryptoProof)Cryptographic proofs are evaluated on correctness, completeness, and rigor: AICrypto Judge Prompt System:You are a professor of theoretical computer science specializing in mathematical foundations of cryptography, grading student homework for CS 6857 (Graduate Cry...

  42. [45]

    Incorrect use of cryptographic definitions (confusing OWF with PRF)

  43. [46]

    Missing verification of critical conditions

  44. [47]

    Intuitive arguments without formal justification

  45. [48]

    Logical gaps in reduction proofs

  46. [49]

    Incorrect probability analysis Output:JSON with score, feedback, key_errors, strengths E.3 Evaluation Summary Table 18 summarizes the evaluation methodology for each benchmark. Benchmark Method Scale Threshold FinanceMath Rule-Based (regex) Binary Exact match OlympiadBench OE Rule-Based (regex) Binaryϵ= 10−6 OlympiadBench TP LLM Judge Binary Correct/Incor...

  47. [2021]

    Training verifiers to solve math word prob- lems.Preprint, arXiv:2110.14168. Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Caio César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo 8 de Rosa, Olli Saarikivi, Adil Salim, Shital Shah, Harkirat Singh Behl, Xin Wang, Sébastien Bubeck, Ronen Eldan, Adam Tauman Kalai, Yi...

    Pith/arXiv arXiv 2023

  48. [2024]

    Association for Computational Linguistics

    Solving for X and beyond: Can large language models solve complex math problems with more- than-two unknowns? InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 16821–16843, Miami, Florida, USA. Association for Computational Linguistics. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yu- taka Matsuo, and Yusuke Iwasawa. 2023. ...

    Pith/arXiv arXiv 2024

  49. [2025]

    In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 35418–35442, Suzhou, China

    Towards robust mathematical reasoning. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 35418–35442, Suzhou, China. Association for Com- putational Linguistics. Iman Mirzadeh, Keivan Alizadeh, Hooman Shahrokhi, Oncel Tuzel, Samy Bengio, and Mehrdad Farajtabar

    2025

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.